RAG means the model fetches live documents before answering. Your visibility therefore depends on two separate things: whether the retriever finds you, and whether the generator chooses to name you.
Almost every practical question in AI visibility resolves to which of those two is failing, and the fixes have nothing in common.
Retrieval-augmented generation is the least glamorous acronym in this field and the most useful thing to understand about it. It is the difference between a system that can only tell you what it memorised and one that goes and looks. That second behaviour is the entire reason AEO is possible: if models only ever recited training data, nothing you did this quarter could change what they say about you.
What problem was RAG built to solve?
Models store knowledge in their weights, and that storage is both frozen and unreliable to query. The paper that named the pattern, Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, opens on exactly that observation: a pre-trained model’s “ability to access and precisely manipulate knowledge is still limited”.
The fix was to stop relying on memory alone. The system keeps the model’s stored knowledge and adds a lookup over an external corpus at query time, so the answer can be assembled from documents rather than recalled.
Fixed at training time and not updatable without retraining. If your company changed after the cutoff, this half of the system does not know.
An external corpus fetched during the request. Live, changeable, and made of documents that exist on the open web today.
What actually happens when someone asks a question?
Asked in natural language, often with constraints the buyer cares about.
Passages scored against the question, top few returned.
Be findableBe self-containedBe crawlableAn answer written from what came back, with sources attached where grounding worked.
Be describableBe agreed uponThe critical property is ordering. Retrieval runs first and it is a hard filter: the generator can only write from what it received. Excellent material that was not retrieved is indistinguishable, from the answer’s point of view, from material that does not exist.
What does the retriever actually match on?
Meaning, increasingly, rather than words. Karpukhin et al. demonstrated dense retrieval, where questions and passages are embedded into vectors and matched by proximity, outperforming a strong term-matching baseline by 9 to 19 points on top-20 passage accuracy.
Two practical consequences follow, and they point in opposite directions from a decade of habit.
If a passage can be matched without containing the query’s words, saying them more often adds little that meaning did not already carry.
Retrieval scores chunks. A section that stands alone is a candidate; one that needs its neighbours to make sense is usually not.
What happens at generation?
The model writes an answer and attaches sources where it can tie a statement to a retrieved document. Anthropic documents this directly: for its web search tool, “citations are always enabled”, and each one carries the URL, the title, and up to 150 characters of the text actually used.
That last detail is worth sitting with. The citation is anchored to a passage, not to a page. The engine is not endorsing your article. It is saying that this specific sentence came from here, which is why a page that answers one thing precisely gets cited more often than one that covers ten things well.
What is a chunk, and why should you care?
A chunk is the passage-sized unit a page gets split into before it is indexed. Engines retrieve chunks, not pages, and that one fact carries more practical weight than anything else in this article.
It means the thing competing for inclusion in an answer is not your carefully structured 2,000-word guide. It is one section of it, judged alone, by a system that never saw the introduction.
Retrieved out of context it says nothing. It cannot be scored well, and if it somehow is, it cannot be quoted.
Complete on its own. Scoreable, quotable, and attributable to you.
The habit this replaces is writing a page as one continuous argument. Continuity is still good for the human reading top to bottom; it just needs to be built out of parts that each stand up alone, rather than out of sentences that depend on the ones before them.
Why do two engines answer differently?
Because each runs its own retrieval over its own corpus. There is no shared index and no shared rulebook.
OpenAI runs OAI-SearchBot for the index behind ChatGPT’s search features, kept separate from its training crawler, and Perplexity documents the same split for PerplexityBot. Google’s AI surfaces draw on the Google index with ordinary Search eligibility. Those are three different corpora assembled by three different crawlers, so disagreement between them is the expected result rather than a bug in your measurement.
If two engines disagree about you, they are not both wrong. They looked in different places and found different things, which is itself a finding about where your evidence lives.
What does this change about what you do?
It gives you a way to tell two failures apart, which is the whole practical payoff.
Retrieval failure
Nothing about you comes back, or the engine describes you vaguely and hedges. Crawler access, missing pages, or content that does not survive being chunked.
Cheap to fix
Generation failure
It describes you well and still does not name you when asked to recommend. The sources it retrieved do not agree about you, or barely mention you at all.
Slow to fix
The mechanics of the second stage, and why consistency beats prominence there, are in how models decide which brands to recommend. If you want to find out which failure you have before reading anything else, the five-minute test settles it.
What does RAG not solve?
Two things, and both matter for how much you should trust an answer about your own company.
It does not guarantee the answer is grounded. Retrieval improves the odds that a claim traces to a document, and the generator can still produce a statement no retrieved source supports. That is why a stale price or a removed feature can appear confidently in an otherwise well-sourced answer, and it is why “the engine cited three sources” is not the same as “everything it said came from them”.
It also does not fix a corpus that is wrong about you. If the documents the retriever finds all repeat an outdated description, retrieval works perfectly and the answer is still wrong. The failure is upstream of the architecture, and the fix is on somebody else’s page rather than in anything a model does.
Both are reasons to read the full answer text rather than a mention count. A tally records that you were named; only the text records that you were named as something you stopped being two years ago.
The acronym is worth remembering for one reason only. RAG is why this is a discipline rather than a lottery: the documents are fetched fresh every time, which means what they say about you is still changeable.